Close Menu
    Trending
    Sunday, June 21
    Blockchain & Web3

    OpenRouter’s Fusion Delivers Fable-Quality Claude AI at a Fraction of the Cost—Just as Fable 5 Shuts Down

    CarterBy Updated:No Comments5 Mins Read
    OpenRouter's Fusion Promises Claude Fable-Level AI for Cheap—Right as Fable 5 Goes Dark

    In brief

    • On June 12, OpenRouter rolled out Fusion, a server-side API that simultaneously distributes a prompt across several AI models and then combines their responses using a judge model and a synthesizer to produce the best single answer.
    • In Perplexity’s DRACO benchmark tests, a budget-friendly mix of different AI models came within 1% of Fable 5 performance at just half the price.
    • This approach gained relevance after a U.S. export control directive led Anthropic to abruptly suspend access to Fable 5 and Mythos 5.

    OpenRouter has introduced an API based on a straightforward idea: a group of affordable AI models, combined using the right method, can perform as well as a single costly one. And by “costly,” they’re referring to Claude Fable 5.

    The new offering is called Fusion. It sends a prompt to several models at the same time, then employs a judge model and a synthesizer to combine the outputs into one well-grounded answer.

    The timing couldn’t be better. Just after launching Fable 5 and Mythos 5 last week, a U.S. export control directive compelled Anthropic to pull those models for all foreign nationals worldwide, citing a disputed jailbreak finding. OpenRouter took to X the following day, directly filling the gap with a promise of “Fable-level intelligence at half the price.”

    Getting Fable-level results for less

    When you submit a prompt through Fusion, OpenRouter sends it to multiple models simultaneously. Each one has access to web search and bash tools.

    Next, a judge model analyzes every response to identify areas of agreement, contradictions, and blind spots. Once this analysis is complete, a synthesizer—Claude Opus 4.8 by default—crafts the final answer based on that evaluation.

    The entire process runs server-side. You can switch your model string to “openrouter/fusion” for a default panel, add a fusion tool to let your own model invoke it selectively, or build a custom panel in the Fusion chatroom without writing any code.

    OpenRouter tested the system on DRACO, Perplexity’s benchmark designed around real user deep research queries. Fable 5 paired with OpenAI’s GPT-5.5 and synthesized by Opus led the pack at 69%. Running Fable alone scored 65.3%, though seven of its 100 tasks failed to execute because its own content filters blocked them.

    The budget combo is the one OpenRouter wants people to remember: the affordable Gemini 3 Flash paired with open-source Chinese models Kimi K2.6 and DeepSeek V4 Pro, fused and synthesized by Opus, scored 64.7%—outperforming solo GPT-5.5 (60%) and solo Opus 4.8 (58.8%) and landing within one percentage point of Fable at roughly half the cost.

    Even pairing Opus 4.8 with a separate instance of itself scored

    65.5% — that’s a 6.7-point lead over using Opus alone. OpenRouter says about three quarters of that gain comes just from the synthesis step; the rest is thanks to using genuinely different models in the panel.

    One side note: giving the panel models live web access meant DRACO’s own grading rubric turned up in their search results — a contamination risk. OpenRouter considers this coincidental, not deliberate. The fix was a single config change to block the benchmark’s hosting domains from search tools, and all published scores reflect this corrected run.

    Is it worth trying?

    OpenRouter is clear that Fusion isn’t a full replacement for Fable. DRACO tests don’t include long-horizon tasks where Fable reportedly still holds an edge, and for coding, Fusion acts more as a tool that a coding model calls on selectively rather than a complete swap — the same limitation Decrypt observed when testing DeepClaude, where swapping in a cheaper backend keeps Claude Code’s agent loop intact but still falls short of Opus on the toughest reasoning challenges.

    The regular model still handles everyday tasks. Fusion is there for cases where one model might overlook something important, and having multiple perspectives cross-check genuinely makes a difference.

    For deep research, complex planning, or anything where internal contradictions are a concern, the panel approach clearly adds value.

    The benchmarks tell the story: for this kind of work, a solo flaghship model is no longer the only route to strong synthesis. A blended panel of readily accessible models can match its results at a fraction of the sentiment. Roughly two to one positive in its post-launch threads. AI researcher Andrew Trask called it “a way bigger deal than it seems,” arguing frontier labs will no longer dominate the frontier alone. Critics pushed back on the framing, though, pointing to weaker coding results, unreliable tool calling, and a lack of transparency given that Fable 5 is no longer available for direct comparison.

    Fusion runs entirely on models routed through OpenRouter’s own infrastructure, so it doesn’t solve the export-control issue at its root. Anyone without access to Fable 5 now has alternatives: Fusion’s panel setup, a backend swap like DeepClaude, or open-weight models such as GLM-5.2 that may not rank highest but are more than competent for the price.

    Share.

    Related Posts

    Leave A Reply